Connectors and MCP

Give the agent Mistral connectors and remote MCP servers.

Guide: Connectors.

def agents.connector(
    name: str,
    *,
    run_as: Literal['auto', 'deployment'] | None = None,
    auto_auth: bool = True,
    credentials_name: str | None = None,
    allow_mcp_ui: bool = False,
) -> Connector

def agents.connector(
    name: str,
    *,
    run_as: Literal['auto', 'deployment'] | None = None,
    auto_auth: bool = True,
    credentials_name: str | None = None,
    allow_mcp_ui: bool = False,
) -> Connector

Give the agent a Mistral connector by name, such as agents.connector("notion").

The connector uses the connection of the user who runs the session. run_as="deployment" uses the connection of the identity the deployment runs with instead; locally, both use your MISTRAL_API_KEY. All connectors of one agent must use the same run_as and auto_auth, and name each connector once. Otherwise the session does not start; agents info and agents serve report the problem before that.

With the default run_as, the agent acts as the user who runs the session, so its deployment must be hardened: only the identities pinned to it can run it. agents serve and agents deploy harden it for you; see Hardened deployments.

A name that matches no connector, or a connector that fails to load, does not stop the session: it starts without that connector, you get an agents.Notice, and the SDK logs a warning such as No Mistral connector is named 'Lienar'. Did you mean 'linear'?.

class Connector(
    connector_name: str,
    *,
    run_as: Literal['auto', 'deployment'] | None = None,
    auto_auth: bool = True,
    credentials_name: str | None = None,
    allow_mcp_ui: bool = False,
)

class Connector(
    connector_name: str,
    *,
    run_as: Literal['auto', 'deployment'] | None = None,
    auto_auth: bool = True,
    credentials_name: str | None = None,
    allow_mcp_ui: bool = False,
)

A Mistral connector that the agent uses, as agents.connector(name) returns it.

def agents.all_connectors(
    *,
    exclude: Sequence[str] = (),
    max_connectors: int | None = None,
    auto_auth: bool = True,
    allow_mcp_ui: bool = False,
) -> AllConnectors

def agents.all_connectors(
    *,
    exclude: Sequence[str] = (),
    max_connectors: int | None = None,
    auto_auth: bool = True,
    allow_mcp_ui: bool = False,
) -> AllConnectors

Give the agent every connector that the user who runs the session has connected.

exclude drops connectors by name. Leave max_connectors unset: it keeps the first connectors by name, not the ones the task needs. Narrow the set with exclude.

auto_auth and allow_mcp_ui apply to every discovered connector and mean what they mean on connector. To give one connector its own credentials_name or allow_mcp_ui, name it with agents.connector(...) in the same list, and the named declaration wins for that connector. It must keep the same auto_auth and the default run_as, or the session does not start. For example, this agent gets every connector but Jira, and Notion with the team credentials:

agents.Agent(
    model="mistral-large-4",
    connectors=[
        agents.all_connectors(exclude=["jira"]),
        agents.connector("notion", credentials_name="team"),
    ]
)

agents.Agent(
    model="mistral-large-4",
    connectors=[
        agents.all_connectors(exclude=["jira"]),
        agents.connector("notion", credentials_name="team"),
    ]
)

class AllConnectors

class AllConnectors

Every connector that the user who runs the session has connected, as of the session's start.

agents.all_connectors() returns it. A connector is in the set when it is active in the user's workspace, has MCP tools, and the user has connected it. With auto_auth, the agent can still find the other connectors and ask the user to connect one.

The set always uses the user's connections: it has no run_as="deployment" form.

class RemoteMCP(
    *,
    url: str,
    auth_header_env_vars: Mapping[str, str] | None = None,
    model_access: ToolModelAccess | None = None,
    disabled_tools: Sequence[str] | None = None,
    startup_timeout_seconds: float | None = None,
    tool_timeout_seconds: float | None = None,
    namespace: str | None = None,
    id: str | None = None,
    name: str | None = None,
)

class RemoteMCP(
    *,
    url: str,
    auth_header_env_vars: Mapping[str, str] | None = None,
    model_access: ToolModelAccess | None = None,
    disabled_tools: Sequence[str] | None = None,
    startup_timeout_seconds: float | None = None,
    tool_timeout_seconds: float | None = None,
    namespace: str | None = None,
    id: str | None = None,
    name: str | None = None,
)

An MCP server the agent reaches over HTTP, declared under a name in Agent(mcps=...).

Each of its tools has the ID <namespace>.<tool name>, such as github.create_issue: permission rules, approvals and hooks name the tool by it. Two servers of one agent cannot share a namespace or an id, so give each its own key in mcps=.

A misspelled keyword raises TypeError, and a bad value raises SdkError naming the argument. Defining the agent raises SdkError for two servers in one namespace, or for a namespace that the built-in tools or the SDK's own tools hold.